Trade-off Analysis by Input Dimensionality Reduction of a BMA Model for Real-Time Battery SOH Estimation in Embedded Environments

Real-time state-of-health (SOH) estimation requires reliable performance with low computational burden. Bayesian model averaging (BMA) provides point estimates and predictive uncertainty, but multiple health indicators and candidate models increase computational cost in resource-constrained embedded systems. In this study, an existing BMA-based SOH estimation method was implemented on a Nucleo-H755ZI-Q board, and the effects of input dimensionality reduction on SOH estimation error, predictive uncertainty, and embedded execution time were analyzed. Nine input-dimensionality conditions were constructed by sequentially excluding health indicators with lower absolute Pearson correlation coefficients. When six input features were used, the mean execution time was 7.266 ms, approximately 19.9 % lower than the 9.070 ms obtained with nine input features, while the RMSE remained comparable to that of the seven-input condition at 0.030. A further reduction from six to five input features shortened the mean execution time by only approximately 1.5 %, whereas the RMSE increased by 46.7 % to 0.044. When four or fewer input features were used, the execution time remained near 6.6 ms, and no substantial additional reduction was observed. These results indicate that six inputs provide a balanced trade-off between SOH estimation performance and execution time under the considered dataset, feature-removal order, and embedded hardware conditions, rather than a generally optimal input dimension.